![\"Albumentation](\"https://i.ibb.co/H4MJWVz/gradcam-album.png\")
For starters I am using these augmentations\n\ntfms = get_transforms(do_flip=True, flip_vert=True, max_rotate=360, max_warp=0, max_zoom=1.1, max_lighting=0.1, p_lighting=0.5,\nxtra_tfms=[(symmetric_warp(magnitude=(-0,0), p=0.5)), rand_crop(p=0.75)] )\n\nI am getting a decent score for every model around 0.72-0.73 but wanted to know what all augmentations can be explored.\nWhat other augmentations worked for others?
For starters I am using these augmentations\n\ntfms = get_transforms(do_flip=True, flip_vert=True, max_rotate=360, max_warp=0, max_zoom=1.1, max_lighting=0.1, p_lighting=0.5,\nxtra_tfms=[(symmetric_warp(magnitude=(-0,0), p=0.5)), rand_crop(p=0.75)] )\n\nI am getting a decent score for every model around 0.72-0.73 but wanted to know what all augmentations can be explored.\nWhat other augmentations worked for others?
@axel81 What's the point of using warp transforms here? Are they really proving to be useful? Because, they seem to particularly useful with Imagenet etc where we can look at the same image from different angles but that doesn't seem to be the case with retina images. Not exactly a transform, but I have a hunch that a good normalization could improve the score!
", "rawMarkdown": "axel81 What's the point of using warp transforms here? Are they really proving to be useful? Because, they seem to particularly useful with Imagenet etc where we can look at the same image from different angles but that doesn't seem to be the case with retina images. Not exactly a transform, but I have a hunch that a good normalization could improve the score!", "votes": null }, { "id": "569950", "postDate": "07/07/2019 15:40:28", "content": "Yes you are right symmetric_warp was not useful, it didn't help me improve anything.
What was your cv score?
", "rawMarkdown": "What was your cv score?", "votes": null }, { "id": "570106", "postDate": "07/07/2019 19:22:57", "content": "hi @axel81 maybe it is time to use fast image augmentation library and easy to use wrapper around other libraries called albumentations
", "rawMarkdown": "hi @axel81 maybe it is time to use fast image augmentation library and easy to use wrapper around other libraries called [albumentations](https://github.com/albu/albumentations)", "votes": null }, { "id": "570267", "postDate": "07/08/2019 04:23:17", "content": "@rinnqd My quadratic_kappa is around 0.891-0.911
", "rawMarkdown": "rinnqd My quadratic_kappa is around 0.891-0.911", "votes": null }, { "id": "570269", "postDate": "07/08/2019 04:23:55", "content": "Do you have any results from that? and what all augmentation would be good to use?
", "rawMarkdown": "Do you have any results from that? and what all augmentation would be good to use?", "votes": null }, { "id": "570938", "postDate": "07/09/2019 00:30:11", "content": "Hi! @axel81, have anyone tried to use some kind of auto-selection of best augmentation techniques, like running a loop that turns them on and off
", "rawMarkdown": "Hi! @axel81, have anyone tried to use some kind of auto-selection of best augmentation techniques, like running a loop that turns them on and off", "votes": null }, { "id": "570941", "postDate": "07/09/2019 00:37:07", "content": "I am using it on pneumothorax competition,i will have to learn more about it and will let you know then,thanks
", "rawMarkdown": "I am using it on pneumothorax competition,i will have to learn more about it and will let you know then,thanks", "votes": null }, { "id": "570967", "postDate": "07/09/2019 01:54:48", "content": "Wow, so complicated, I can't comprehend it.
", "rawMarkdown": "Wow, so complicated, I can't comprehend it.", "votes": null }, { "id": "571382", "postDate": "07/09/2019 14:46:07", "content": "I haven't
", "rawMarkdown": "I haven't", "votes": null }, { "id": "572200", "postDate": "07/10/2019 15:55:49", "content": "We use almost the same, but I think the unbalanced number in different target type should be improved.
", "rawMarkdown": "We use almost the same, but I think the unbalanced number in different target type should be improved.", "votes": null }, { "id": "573383", "postDate": "07/12/2019 07:38:19", "content": "I illustrate 6 sensible data augmentation with visualization by Grad CAM here in this kernel. Hope it is helpful.
\n\n
@axel81 What's the point of using warp transforms here? Are they really proving to be useful? Because, they seem to particularly useful with Imagenet etc where we can look at the same image from different angles but that doesn't seem to be the case with retina images. Not exactly a transform, but I have a hunch that a good normalization could improve the score!
", "votes": null, "replies": [ { "id": 569950, "author_name": "axel81", "author_url": "", "post_date": "07/07/2019 15:40:28", "content": "Yes you are right symmetric_warp was not useful, it didn't help me improve anything.
What was your cv score?
", "votes": null, "replies": [ { "id": 570267, "author_name": "axel81", "author_url": "", "post_date": "07/08/2019 04:23:17", "content": "@rinnqd My quadratic_kappa is around 0.891-0.911
", "votes": null, "replies": [] } ] }, { "id": 570106, "author_name": "mobassir", "author_url": "", "post_date": "07/07/2019 19:22:57", "content": "hi @axel81 maybe it is time to use fast image augmentation library and easy to use wrapper around other libraries called albumentations
", "votes": null, "replies": [ { "id": 570269, "author_name": "axel81", "author_url": "", "post_date": "07/08/2019 04:23:55", "content": "Do you have any results from that? and what all augmentation would be good to use?
", "votes": null, "replies": [] }, { "id": 570941, "author_name": "mobassir", "author_url": "", "post_date": "07/09/2019 00:37:07", "content": "I am using it on pneumothorax competition,i will have to learn more about it and will let you know then,thanks
", "votes": null, "replies": [] } ] }, { "id": 570938, "author_name": "cv13j0", "author_url": "", "post_date": "07/09/2019 00:30:11", "content": "Hi! @axel81, have anyone tried to use some kind of auto-selection of best augmentation techniques, like running a loop that turns them on and off
", "votes": null, "replies": [ { "id": 571382, "author_name": "axel81", "author_url": "", "post_date": "07/09/2019 14:46:07", "content": "I haven't
", "votes": null, "replies": [] } ] }, { "id": 570967, "author_name": "kidsclothing", "author_url": "", "post_date": "07/09/2019 01:54:48", "content": "Wow, so complicated, I can't comprehend it.
", "votes": null, "replies": [] }, { "id": 572200, "author_name": "xujingzhao", "author_url": "", "post_date": "07/10/2019 15:55:49", "content": "We use almost the same, but I think the unbalanced number in different target type should be improved.
", "votes": null, "replies": [] }, { "id": 573383, "author_name": "ratthachat", "author_url": "", "post_date": "07/12/2019 07:38:19", "content": "I illustrate 6 sensible data augmentation with visualization by Grad CAM here in this kernel. Hope it is helpful.
\n\n